Computation Offloading and Resource Allocation Based on P-DQN in LEO Satellite Edge Networks

Area of Science:

• Satellite Communications
• Edge Computing
• Network Resource Management

Background:

• Traditional Low Earth Orbit (LEO) satellite networks face capacity limitations and are often independent of terrestrial networks.
• Integrating Mobile Edge Computing (MEC) with LEO networks creates an "end-edge-cloud" architecture for enhanced task offloading.
• Dynamic network conditions and complex decision-making pose challenges for LEO satellite edge networks.

Purpose of the Study:

• To develop a method for joint computation offloading and resource allocation in dynamic LEO satellite edge networks.
• To address the challenges of discrete-continuous hybrid action spaces and time-varying network dynamics.
• To maximize the long-term number of satisfied tasks in the integrated network.

Main Methods:

• Modeling time-varying channel characteristics.
• Constructing communication and computation models for three offloading scenarios.
• Formulating constraints for task offloading, resource availability, and power control.
• Utilizing Parameterized Deep Q-Network (P-DQN) and Parameterized Action Markov Decision Process (PAMDP) for real-time decision-making.

Main Results:

• The proposed P-DQN method demonstrates effectiveness in joint computation offloading, resource allocation, and power control.
• Simulation results show the P-DQN approach approaches optimal control in dynamic LEO satellite edge networks.
• P-DQN outperforms other reinforcement learning methods designed for single action spaces (discrete or continuous).

Conclusions:

• The P-DQN approach offers a robust solution for optimizing task satisfaction in integrated LEO satellite and MEC environments.
• This method successfully handles the complexities of hybrid action spaces and network dynamics.
• The study highlights the potential of advanced reinforcement learning techniques for future satellite edge network management.